Qwen4-Max
by Together
Today's price · per 1M tokens
Input
$1.05
per 1M tokens
Output
$3.15
per 1M tokens
Blended
$1.58
blended $/1M — 3:1 weighted input:output
Cached input
$0.075
7% of input — prompt caching
How this price is scoped: Together's pricing page lists Qwen4 Max at $1.50 input / $0.075 cached input / $4.50 output per 1M tokens, with a promo price of $1.05 input / $3.15 output valid until 11/23/2026. We publish the rate you would pay today and re-verify it before the changeover. Together's serverless models list printed $1.425 / $0.07125 / $4.275 on 2026-11-18 and 2026-11-19; the pricing page is the rate card this row cites.
Price receipt
We read Together's own pricing page on and found Qwen4 Max listed with a price on the same row — that read is where the number above comes from, recorded as a price change. We keep the page text we read; its content hash is 6bc57ac999.
Overview
Alibaba's next-generation Qwen flagship, served by Together on its serverless pay-per-token API. It takes text and image input and returns text, with function calling, structured outputs and a 1M-token context. Together lists it at $1.50/$4.50 per 1M tokens, with a launch promo of $1.05/$3.15 until November 23, 2026.
Capabilities
struck through = not supportedBenchmark performance
accuracy % · higher is betterNo per-benchmark accuracy scores are shown for this model. Qwen's launch post (read 2026-11-21) reports Qwen4-Max on agent and work benchmarks we do not track, Terminal-Bench 4.0 (we track 2.1), and Humanity's Last Exam without saying whether tools were used, so none fits a column here. Vals and Vellum list no independent score for a benchmark we track, and Epoch AI has not run it. The percentile above is its standing across the independent composites below.
- AA Intelligence Index: 42.9 (Artificial Analysis composite across reasoning, knowledge and coding evals)
- AA-Omniscience Index: 27.7 (−100–100)
- GDPval-AA v2: 1808.7 ELO (Real-world work tasks, human baseline = 1000)
- HLE-Diamond: 44.9% (measured 2026-11-22) — A revised 1,000-question set of Humanity's Last Exam, half reasoning, half knowledge
- CritPt v2, hard subset: 43% (measured 2026-11-22) — The hardest research-level physics problems in CritPt v2, solved with a code sandbox
- FrontierCode-AA 2.0: 26.5% (measured 2026-11-22) — Coding tasks in real open-source repositories, graded on whether a maintainer would merge the change
- Terminal-Bench-Science: 23.3% (measured 2026-11-22) — Expert-written science research workflows completed in a terminal
- GDP.pdf: 26.2% (measured 2026-11-22) — Tasks over long professional documents; share of tasks that pass every criterion
- AutomationBench-AA: 72.6% (measured 2026-11-22) — Multi-step work across simulated business apps, without breaking the rules set
- MLCR-AA: 19.4% (measured 2026-11-22) — Reasoning over long, fragmented medical records; pass rate
- Terminal-Bench 4.0: 48.5% (measured 2026-11-22)
- Ranks #19 of 207 comparably-measured models by percentile score, across 3 independent measurements
- Ranks #53 of 207 comparably-tested models by normalized performance per dollar
Source: Artificial Analysis · updated · See full rankings →
Specifications
- Provider
- Together
- Context window
- 1M tokens
- Modality
- text, image
- Parameters
- Proprietary
- Open source
- No — proprietary
- Released
- Status
- Current
- Last updated
- Tags
Availability verified: — listed on Together's own page